NCCL深度学习之初始化及ncclUniqueId的产生源码解析

Rose ·

更新时间:2024-11-14

· 887 次阅读

作者｜KIDGINBROOK

更新｜潘丽晨

NCCL是英伟达开源的GPU通信库，支持集合通信和点对点通信。

看下官方给的一个demo：

#include &lt;stdio.h&gt;
#include "cuda_runtime.h"
#include "nccl.h"
#include "mpi.h"
#include &lt;unistd.h&gt;
#include &lt;stdint.h&gt;
#define MPICHECK(cmd) do {                          \
  int e = cmd;                                      \
  if( e != MPI_SUCCESS ) {                          \
    printf("Failed: MPI error %s:%d '%d'\n",        \
        __FILE__,__LINE__, e);   \
    exit(EXIT_FAILURE);                             \
  }                                                 \
} while(0)
#define CUDACHECK(cmd) do {                         \
  cudaError_t e = cmd;                              \
  if( e != cudaSuccess ) {                          \
    printf("Failed: Cuda error %s:%d '%s'\n",             \
        __FILE__,__LINE__,cudaGetErrorString(e));   \
    exit(EXIT_FAILURE);                             \
  }                                                 \
} while(0)
#define NCCLCHECK(cmd) do {                         \
  ncclResult_t r = cmd;                             \
  if (r!= ncclSuccess) {                            \
    printf("Failed, NCCL error %s:%d '%s'\n",             \
        __FILE__,__LINE__,ncclGetErrorString(r));   \
    exit(EXIT_FAILURE);                             \
  }                                                 \
} while(0)
static uint64_t getHostHash(const char* string) {
  // Based on DJB2a, result = result * 33 ^ char
  uint64_t result = 5381;
  for (int c = 0; string[c] != '\0'; c++){
    result = ((result &lt;&lt; 5) + result) ^ string[c];
  }
  return result;
}
static void getHostName(char* hostname, int maxlen) {
  gethostname(hostname, maxlen);
  for (int i=0; i&lt; maxlen; i++) {
    if (hostname[i] == '.') {
        hostname[i] = '\0';
        return;
    }
  }
}
int main(int argc, char* argv[])
{
  int size = 32*1024*1024;
  int myRank, nRanks, localRank = 0;
  //initializing MPI
  MPICHECK(MPI_Init(&amp;argc, &amp;argv));
  MPICHECK(MPI_Comm_rank(MPI_COMM_WORLD, &amp;myRank));
  MPICHECK(MPI_Comm_size(MPI_COMM_WORLD, &amp;nRanks));
  //calculating localRank which is used in selecting a GPU
  uint64_t hostHashs[nRanks];
  char hostname[1024];
  getHostName(hostname, 1024);
  hostHashs[myRank] = getHostHash(hostname);
  MPICHECK(MPI_Allgather(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, hostHashs, sizeof(uint64_t), MPI_BYTE, MPI_COMM_WORLD));
  for (int p=0; p&lt;nRanks; p++) {
     if (p == myRank) break;
     if (hostHashs[p] == hostHashs[myRank]) localRank++;
  }
  //each process is using two GPUs
  int nDev = 2;
  float** sendbuff = (float**)malloc(nDev * sizeof(float*));
  float** recvbuff = (float**)malloc(nDev * sizeof(float*));
  cudaStream_t* s = (cudaStream_t*)malloc(sizeof(cudaStream_t)*nDev);
  //picking GPUs based on localRank
  for (int i = 0; i &lt; nDev; ++i) {
    CUDACHECK(cudaSetDevice(localRank*nDev + i));
    CUDACHECK(cudaMalloc(sendbuff + i, size * sizeof(float)));
    CUDACHECK(cudaMalloc(recvbuff + i, size * sizeof(float)));
    CUDACHECK(cudaMemset(sendbuff[i], 1, size * sizeof(float)));
    CUDACHECK(cudaMemset(recvbuff[i], 0, size * sizeof(float)));
    CUDACHECK(cudaStreamCreate(s+i));
  }
  ncclUniqueId id;
  ncclComm_t comms[nDev];
  //generating NCCL unique ID at one process and broadcasting it to all
  if (myRank == 0) ncclGetUniqueId(&amp;id);
  MPICHECK(MPI_Bcast((void *)&amp;id, sizeof(id), MPI_BYTE, 0, MPI_COMM_WORLD));
  //initializing NCCL, group API is required around ncclCommInitRank as it is
  //called across multiple GPUs in each thread/process
  NCCLCHECK(ncclGroupStart());
  for (int i=0; i&lt;nDev; i++) {
     CUDACHECK(cudaSetDevice(localRank*nDev + i));
     NCCLCHECK(ncclCommInitRank(comms+i, nRanks*nDev, id, myRank*nDev + i));
  }
  NCCLCHECK(ncclGroupEnd());
  //calling NCCL communication API. Group API is required when using
  //multiple devices per thread/process
  NCCLCHECK(ncclGroupStart());
  for (int i=0; i&lt;nDev; i++)
     NCCLCHECK(ncclAllReduce((const void*)sendbuff[i], (void*)recvbuff[i], size, ncclFloat, ncclSum,
           comms[i], s[i]));
  NCCLCHECK(ncclGroupEnd());
  //synchronizing on CUDA stream to complete NCCL communication
  for (int i=0; i&lt;nDev; i++)
      CUDACHECK(cudaStreamSynchronize(s[i]));
  //freeing device memory
  for (int i=0; i&lt;nDev; i++) {
     CUDACHECK(cudaFree(sendbuff[i]));
     CUDACHECK(cudaFree(recvbuff[i]));
  }
  //finalizing NCCL
  for (int i=0; i&lt;nDev; i++) {
     ncclCommDestroy(comms[i]);
  }
  //finalizing MPI
  MPICHECK(MPI_Finalize());
  printf("[MPI Rank %d] Success \n", myRank);
  return 0;
}

在上边的示例中，rank0会执行ncclGetUniqueId获取Id，然后通过mpi广播给其他rank，接下来看下UniqueId是怎么产生的。

ncclResult_t ncclGetUniqueId(ncclUniqueId* out) {
  NCCLCHECK(ncclInit());
  NCCLCHECK(PtrCheck(out, "GetUniqueId", "out"));
  return bootstrapGetUniqueId(out);
}

然后看下ncclInit。

首先执行initEnv，设置环境变量。

然后执行initNet，用来初始化nccl所需要的网络，包括两个，一个是bootstrap网络，另外一个是数据通信网络，bootstrap网络主要用于初始化时交换一些简单的信息，比如每个机器的ip端口，由于数据量很小，而且主要是在初始化阶段执行一次，因此bootstrap使用的是tcp；而通信网络是用于实际数据的传输，因此会优先使用rdma（支持gdr的话会优先使用gdr）。

ncclResult_t initNet() {
  // Always initialize bootstrap network
  NCCLCHECK(bootstrapNetInit());
  NCCLCHECK(initNetPlugin(&amp;ncclNet, &amp;ncclCollNet));
  if (ncclNet != NULL) return ncclSuccess;
  if (initNet(&amp;ncclNetIb) == ncclSuccess) {
    ncclNet = &amp;ncclNetIb;
  } else {
    NCCLCHECK(initNet(&amp;ncclNetSocket));
    ncclNet = &amp;ncclNetSocket;
  }
  return ncclSuccess;
}

bootstrapNetInit就是bootstrap网络的初始化，主要就是通过findInterfaces遍历机器上所有的网卡信息，通过prefixList匹配选择使用哪些网卡，将可用网卡的信息保存下来，将ifa_name保存到全局的bootstrapNetIfNames，ip地址保存到全局bootstrapNetIfAddrs，默认除了docker和lo其他的网卡都可以使用。

例如在测试机器上有三张网卡，分别是xgbe0、xgbe1、xgbe2，那么就会把这三个ifaname和对应的ip地址保存下来，另外nccl提供了环境变量NCCL_SOCKET_IFNAME可以用来指定想用的网卡名，例如通过export NCCL_SOCKET_IFNAME=xgbe0来指定使用xgbe0，其实就是通过prefixList来匹配做到的。

static int findInterfaces(const char* prefixList, char* names, union socketAddress *addrs, int sock_family, int maxIfNameSize, int maxIfs) {
  struct netIf userIfs[MAX_IFS];
  bool searchNot = prefixList &amp;&amp; prefixList[0] == '^';
  if (searchNot) prefixList++;
  bool searchExact = prefixList &amp;&amp; prefixList[0] == '=';
  if (searchExact) prefixList++;
  int nUserIfs = parseStringList(prefixList, userIfs, MAX_IFS);
  int found = 0;
  struct ifaddrs *interfaces, *interface;
  getifaddrs(&amp;interfaces);
  for (interface = interfaces; interface &amp;&amp; found &lt; maxIfs; interface = interface-&gt;ifa_next) {
    if (interface-&gt;ifa_addr == NULL) continue;
    int family = interface-&gt;ifa_addr-&gt;sa_family;
    if (family != AF_INET &amp;&amp; family != AF_INET6)
      continue;
    if (sock_family != -1 &amp;&amp; family != sock_family)
      continue;
    if (family == AF_INET6) {
      struct sockaddr_in6* sa = (struct sockaddr_in6*)(interface-&gt;ifa_addr);
      if (IN6_IS_ADDR_LOOPBACK(&amp;sa-&gt;sin6_addr)) continue;
    }
    if (!(matchIfList(interface-&gt;ifa_name, -1, userIfs, nUserIfs, searchExact) ^ searchNot)) {
      continue;
    }
    bool duplicate = false;
    for (int i = 0; i &lt; found; i++) {
      if (strcmp(interface-&gt;ifa_name, names+i*maxIfNameSize) == 0) { duplicate = true; break; }
    }
    if (!duplicate) {
      strncpy(names+found*maxIfNameSize, interface-&gt;ifa_name, maxIfNameSize);
      int salen = (family == AF_INET) ? sizeof(sockaddr_in) : sizeof(sockaddr_in6);
      memcpy(addrs+found, interface-&gt;ifa_addr, salen);
      found++;
    }
  }
  freeifaddrs(interfaces);
  return found;
}

开始初始化通信网络。

ncclNet_t结构体是一系列的函数指针，比如初始化，发送，接收等；socket，IB等通信方式都实现了自己的ncclNet_t，如ncclNetSocket，ncclNetIb，初始化通信网络的过程就是依次看哪个通信模式可用，然后赋值给全局的ncclNet。

首先执行initNetPlugin，查看是否有libnccl-net.so，测试环境没有这个so，所以直接返回。

然后尝试使用IB网络：

首先执行ncclNetIb的init函数，就是ncclIbInit。

ncclResult_t ncclIbInit(ncclDebugLogger_t logFunction) {
  static int shownIbHcaEnv = 0;
  if(wrap_ibv_symbols() != ncclSuccess) { return ncclInternalError; }
  if (ncclParamIbDisable()) return ncclInternalError;
  if (ncclNIbDevs == -1) {
    pthread_mutex_lock(&amp;ncclIbLock);
    wrap_ibv_fork_init();
    if (ncclNIbDevs == -1) {
      ncclNIbDevs = 0;
      if (findInterfaces(ncclIbIfName, &amp;ncclIbIfAddr, MAX_IF_NAME_SIZE, 1) != 1) {
        WARN("NET/IB : No IP interface found.");
        return ncclInternalError;
      }
      // Detect IB cards
      int nIbDevs;
      struct ibv_device** devices;
      // Check if user defined which IB device:port to use
      char* userIbEnv = getenv("NCCL_IB_HCA");
      if (userIbEnv != NULL &amp;&amp; shownIbHcaEnv++ == 0) INFO(NCCL_NET|NCCL_ENV, "NCCL_IB_HCA set to %s", userIbEnv);
      struct netIf userIfs[MAX_IB_DEVS];
      bool searchNot = userIbEnv &amp;&amp; userIbEnv[0] == '^';
      if (searchNot) userIbEnv++;
      bool searchExact = userIbEnv &amp;&amp; userIbEnv[0] == '=';
      if (searchExact) userIbEnv++;
      int nUserIfs = parseStringList(userIbEnv, userIfs, MAX_IB_DEVS);
      if (ncclSuccess != wrap_ibv_get_device_list(&amp;devices, &amp;nIbDevs)) return ncclInternalError;
      for (int d=0; d&lt;nIbDevs &amp;&amp; ncclNIbDevs&lt;MAX_IB_DEVS; d++) {
        struct ibv_context * context;
        if (ncclSuccess != wrap_ibv_open_device(&amp;context, devices[d]) || context == NULL) {
          WARN("NET/IB : Unable to open device %s", devices[d]-&gt;name);
          continue;
        }
        int nPorts = 0;
        struct ibv_device_attr devAttr;
        memset(&amp;devAttr, 0, sizeof(devAttr));
        if (ncclSuccess != wrap_ibv_query_device(context, &amp;devAttr)) {
          WARN("NET/IB : Unable to query device %s", devices[d]-&gt;name);
          if (ncclSuccess != wrap_ibv_close_device(context)) { return ncclInternalError; }
          continue;
        }
        for (int port = 1; port &lt;= devAttr.phys_port_cnt; port++) {
          struct ibv_port_attr portAttr;
          if (ncclSuccess != wrap_ibv_query_port(context, port, &amp;portAttr)) {
            WARN("NET/IB : Unable to query port %d", port);
            continue;
          }
          if (portAttr.state != IBV_PORT_ACTIVE) continue;
          if (portAttr.link_layer != IBV_LINK_LAYER_INFINIBAND
              &amp;&amp; portAttr.link_layer != IBV_LINK_LAYER_ETHERNET) continue;
          // check against user specified HCAs/ports
          if (! (matchIfList(devices[d]-&gt;name, port, userIfs, nUserIfs, searchExact) ^ searchNot)) {
            continue;
          }
          TRACE(NCCL_INIT|NCCL_NET,"NET/IB: [%d] %s:%d/%s ", d, devices[d]-&gt;name, port,
              portAttr.link_layer == IBV_LINK_LAYER_INFINIBAND ? "IB" : "RoCE");
          ncclIbDevs[ncclNIbDevs].device = d;
          ncclIbDevs[ncclNIbDevs].guid = devAttr.sys_image_guid;
          ncclIbDevs[ncclNIbDevs].port = port;
          ncclIbDevs[ncclNIbDevs].link = portAttr.link_layer;
          ncclIbDevs[ncclNIbDevs].speed = ncclIbSpeed(portAttr.active_speed) * ncclIbWidth(portAttr.active_width);
          ncclIbDevs[ncclNIbDevs].context = context;
          strncpy(ncclIbDevs[ncclNIbDevs].devName, devices[d]-&gt;name, MAXNAMESIZE);
          NCCLCHECK(ncclIbGetPciPath(ncclIbDevs[ncclNIbDevs].devName, &amp;ncclIbDevs[ncclNIbDevs].pciPath, &amp;ncclIbDevs[ncclNIbDevs].realPort));
          ncclIbDevs[ncclNIbDevs].maxQp = devAttr.max_qp;
          ncclNIbDevs++;
          nPorts++;
          pthread_create(&amp;ncclIbAsyncThread, NULL, ncclIbAsyncThreadMain, context);
        }
        if (nPorts == 0 &amp;&amp; ncclSuccess != wrap_ibv_close_device(context)) { return ncclInternalError; }
      }
      if (nIbDevs &amp;&amp; (ncclSuccess != wrap_ibv_free_device_list(devices))) { return ncclInternalError; };
    }
    if (ncclNIbDevs == 0) {
      INFO(NCCL_INIT|NCCL_NET, "NET/IB : No device found.");
    } else {
      char line[1024];
      line[0] = '\0';
      for (int d=0; d&lt;ncclNIbDevs; d++) {
        snprintf(line+strlen(line), 1023-strlen(line), " [%d]%s:%d/%s", d, ncclIbDevs[d].devName,
            ncclIbDevs[d].port, ncclIbDevs[d].link == IBV_LINK_LAYER_INFINIBAND ? "IB" : "RoCE");
      }
      line[1023] = '\0';
      char addrline[1024];
      INFO(NCCL_INIT|NCCL_NET, "NET/IB : Using%s ; OOB %s:%s", line, ncclIbIfName, socketToString(&amp;ncclIbIfAddr.sa, addrline));
    }
    pthread_mutex_unlock(&amp;ncclIbLock);
  }
  return ncclSuccess;
}

首先第三行通过wrap_ibv_symbols加载动态库libibverbs.so，然后获取动态库的各个函数。

然后通过wrap_ibv_fork_init避免fork引起rdma网卡读写出错。

后面会讲到ib网络也会用到socket进行带外网络的传输，所以这里也通过findInterfaces获取一个可用的网卡保存到ncclIbIfAddr。

通过ibv_get_device_list获取所有rdma设备到devices中，遍历devices的每个device，因为每个HCA可能有多个物理port，所以对每个device遍历每一个物理port，获取每个port的信息。

然后将相关信息保存到全局的ncclIbDevs中，比如是哪个device的哪个port，使用的是IB还是ROCE，device的pci路径，maxqp，device的name等，注意这里也有类似bootstrap网络NCCL_SOCKET_IFNAME的环境变量，叫NCCL_IB_HCA，可以指定使用哪个IB HCA。

到这里整个初始化的过程就完成了，一句话总结就是，获取了当前机器上所有可用的IB网卡和普通以太网卡之后保存下来。

然后开始生成UniqueId。

ncclResult_t bootstrapCreateRoot(ncclUniqueId* id, bool idFromEnv) {
  ncclNetHandle_t* netHandle = (ncclNetHandle_t*) id;
  void* listenComm;
  NCCLCHECK(bootstrapNetListen(idFromEnv ? dontCareIf : 0, netHandle, &amp;listenComm));
  pthread_t thread;
  pthread_create(&amp;thread, NULL, bootstrapRoot, listenComm);
  return ncclSuccess;
}

ncclNetHandle_t也是一个字符数组，然后执行bootstrapNetListen。

static ncclResult_t bootstrapNetListen(int dev, ncclNetHandle_t* netHandle, void** listenComm) {
  union socketAddress* connectAddr = (union socketAddress*) netHandle;
  static_assert(sizeof(union socketAddress) &lt; NCCL_NET_HANDLE_MAXSIZE, "union socketAddress size is too large");
  // if dev &gt;= 0, listen based on dev
  if (dev &gt;= 0) {
    NCCLCHECK(bootstrapNetGetSocketAddr(dev, connectAddr));
  } else if (dev == findSubnetIf) {
    ...
  } // Otherwise, handle stores a local address
  struct bootstrapNetComm* comm;
  NCCLCHECK(bootstrapNetNewComm(&amp;comm));
  NCCLCHECK(createListenSocket(&amp;comm-&gt;fd, connectAddr));
  *listenComm = comm;
  return ncclSuccess;
}

依次看下这三个函数，通过bootstrapNetGetSocketAddr获取一个可用的ip地址。

static ncclResult_t bootstrapNetGetSocketAddr(int dev, union socketAddress* addr) {
  if (dev &gt;= bootstrapNetIfs) return ncclInternalError;
  memcpy(addr, bootstrapNetIfAddrs+dev, sizeof(*addr));
  return ncclSuccess;
}

此时dev是0， bootstrapNetIfs是初始化bootstrap网络的时候一共找到了几个可用的网卡，这里就是获取了第0个可用的ip地址。

然后通过bootstrapNetNewComm创建bootstrapNetComm，bootstrapNetComm其实就是fd，bootstrapNetNewComm其实就是new了一个bootstrapNetComm。

struct bootstrapNetComm {
  int fd; 
};

通过createListenSocket启动socker server。

static ncclResult_t createListenSocket(int *fd, union socketAddress *localAddr) {
  /* IPv4/IPv6 support */
  int family = localAddr->sa.sa_family;
  int salen = (family == AF_INET) ? sizeof(sockaddr_in) : sizeof(sockaddr_in6);
  /* Create socket and bind it to a port */
  int sockfd = socket(family, SOCK_STREAM, 0);
  if (sockfd == -1) {
    WARN("Net : Socket creation failed : %s", strerror(errno));
    return ncclSystemError;
  }
  if (socketToPort(&localAddr->sa)) {
    // Port is forced by env. Make sure we get the port.
    int opt = 1;
#if defined(SO_REUSEPORT)
    SYSCHECK(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR | SO_REUSEPORT, &opt, sizeof(opt)), "setsockopt");
#else
    SYSCHECK(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)), "setsockopt");
#endif
  }
  // localAddr port should be 0 (Any port)
  SYSCHECK(bind(sockfd, &localAddr->sa, salen), "bind");
  /* Get the assigned Port */
  socklen_t size = salen;
  SYSCHECK(getsockname(sockfd, &localAddr->sa, &size), "getsockname");
#ifdef ENABLE_TRACE
  char line[1024];
  TRACE(NCCL_INIT|NCCL_NET,"Listening on socket %s", socketToString(&localAddr->sa, line));
#endif
  /* Put the socket in listen mode
   * NB: The backlog will be silently truncated to the value in /proc/sys/net/core/somaxconn
   */
  SYSCHECK(listen(sockfd, 16384), "listen");
  *fd = sockfd;
  return ncclSuccess;
}

创建监听fd，ip由localaddr指定，初始端口为0，bind时随机找一个可用端口，并通过getsockname(sockfd, &localAddr->sa, &size)将ip端口写回到localaddr，这里localaddr就是UniqueId。

到这里UniqueId也就产生了，其实就是当前机器的ip和port。

欢迎 Star、试用 OneFlow 最新版本：github.com/Oneflow-Inc…

以上就是NCCL源码解析之初始化及ncclUniqueId的产生详解的详细内容，更多关于NCCL初始化ncclUniqueId产生的资料请关注软件开发网其它相关文章！

学习源码初始化

1024 个赞